Researchers Want to Police AI at Work — In Under a Second

AI is now inside most workplaces — 78% of organizations worldwide, according to the paper — but the rules for controlling it haven't kept up. The authors describe an 'attestation deficit': a company may have a policy saying 'don't feed customer data into chatbots,' yet when a regulator asks for proof that the policy is enforced, nobody can produce it. This isn't a technology problem, they argue. It's an organizational and design problem, and it's getting worse as AI spreads.

The paper pulls together some uncomfortable numbers. The Stanford 2026 AI Index counts 362 documented AI incidents. IBM's breach study puts the average data breach at $4.99 million, with 92% of victims lacking basic access controls. And a survey found only 38% of companies monitor AI end-to-end, with just 17% watching what happens when one AI agent (software that takes actions on your behalf) talks to another. That last gap is the scary one: unsupervised AI talking to unsupervised AI.

Their proposed answer is AGIL, a five-layer system. Think of it like a security desk for AI. Layer one hunts for 'shadow AI' — employees quietly using tools nobody approved. Layer two scores behavior for security, made-up answers (hallucinations), privacy, and accountability. Layer three is a gateway that permits, blocks, or modifies an AI action in under 100 milliseconds — faster than a human blink. Layer four writes tamper-proof audit records automatically, as a side effect of doing its job. Layer five uses machine learning to update the rules as laws change across countries.

The catch: this is a theoretical framework. The authors openly say it hasn't been tested in a real deployment — no pilot, no measured results, just architecture and 20 references. There's also an irony worth noting: fixing AI oversight with more AI adds another automated system to watch. Until someone builds and tests it, 'governing at machine speed' remains a proposal, not a product you can buy.